henhaobio
Sign in

bioRxivBioengineeringPreprint

BioPathfinder: Evidence-guided multi-agent platform enables hypothesis discovery for CAR-T engineering

Shubing Wang, Yan-Ruide LiiD, Qian Wang, Youcheng Yang, Xinyuan SheniD, Haoran Li, Haochen NaniD, Z J Chen, Yichen ZhuiD, B Y ZhangiD, Haoyan Ding, Jennifer SotoiD, Soomin Park, Yi ZhengiD, Xiangliang Huang, Debiao LiiD

PDFDOICode
In-depth full-text digest available · Jump to digest ↓

Abstract

Clinical studies of chimeric antigen receptor (CAR)-T therapy generate diverse molecular and clinical evidence that remains fragmented across publications, public repositories and patient-derived datasets, limiting systematic therapeutic discovery. Here we present BioPathfinder, an evidence-guided multi-agent workflow for closed-loop biomedical discovery. BioPathfinder constructs a provenance-aware knowledge resource linking publications with patient single-cell RNA sequencing (scRNA-seq) datasets and clinical metadata, and uses role-specialized large language model agents to generate, review and prioritize diverse, falsifiable and dataset-aware mechanistic hypotheses for computational and experimental validation. Applied to a curated corpus of CAR-T-treated patient studies and matched scRNA-seq datasets, BioPathfinder identified candidate mechanisms underlying CAR-T persistence, dysfunction and therapeutic resistance. The workflow prioritized the hypothesis that genes associated with an NK-like transition programme could be targeted to reduce CAR-T exhaustion and improve persistence. Analysis of patient scRNA-seq datasets showed enrichment of this programme in exhausted post-infusion CAR-T cells. Virtual perturbation prioritized transition-associated receptor genes, including KLRC1, KLRD1 and KLRG1, and expert review selected KLRC1, encoding NKG2A, for experimental validation. In vitro and in vivo chronic-stimulation models showed that NKG2A marked activated, exhaustion-associated CD8 CAR-T cells, whereas NKG2A blockade enhanced antitumour activity and persistence-associated functional readouts in vivo. BioPathfinder establishes a generalizable framework that transforms fragmented clinical single-cell evidence into experimentally validated therapeutic hypotheses, providing a scalable strategy for AI-guided biomedical discovery. HIGHLIGHTSO_LIBioPathfinder integrates fragmented CAR-T clinical studies, patient scRNA-seq datasets and metadata into a provenance-aware evidence resource for AI-guided hypothesis discovery. C_LIO_LIA multi-agent workflow generates, critiques and prioritizes diverse, falsifiable, dataset-aware mechanisms underlying CAR-T persistence, dysfunction and therapeutic resistance. C_LIO_LIBioPathfinder prioritizes KLRC1/NKG2A as a therapeutic target, and experimental validation demonstrates that NKG2A blockade enhances CAR-T persistence and antitumour function. C_LI

In-depth summary

Claude Sonnet 4.6 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (28 Jul 2026) · CC BY 4.0

A provenance-tracked multi-agent system curated 57 CAR-T patient scRNA-seq studies, generated reviewable hypotheses, and prioritized KLRC1/NKG2A, whose blockade improved CD8+ CAR-T antitumour activity in a Raji xenograft model.

Evidence on CAR-T persistence and dysfunction is scattered across papers, supplements and repositories, and a claim is only interpretable in its original sampling and treatment context. Document retrieval alone does not say which cohort, accession or cell population supports a mechanism; this work builds a provenance-tracked paper-dataset resource and feeds it to planner and reviewer agents.

Key results

  1. ShownThe Curator agent narrowed a large literature pool to a frozen core of dataset-linked CAR-T patient scRNA-seq studies, recovering more studies than a literature-derived list or two broad single-cell resources.

    16,597 records led to 69 candidates; 57 strictly eligible, 10 supporting only, 2 excluded. Curator found 57 core studies vs 39 (literature list), 2 (scBaseCount) and 3 (CZ CELLxGENE). Against expert ground truth: recall 90.5%, precision 82.6%, F1 86.4%. · BioPathfinder constructs a structured evidence…; Curator evidence extraction and database…

  2. ShownAn agent-assisted reviewer (curated context plus network checks) scored hypotheses more repeatably than a rubric-only LLM baseline, using GPT-5.4-mini for both, and gave a less compressed score distribution.

    50 hypotheses scored in three repeated runs; the agent route had higher pairwise Spearman correlation and ICC(3,1) than the baseline; paired hypothesis-level bootstrap on Fisher-z Spearman, P = 0.016. Mean correlation values are not given in the text. · BioPathfinder generates and prioritizes…; Reviewer routing and hypothesis-merit…; Figure legends

  3. SuggestedIn the patient dataset GSE125881, NK-like transition programmes rose after infusion and tracked Temra-like states and pseudotime. Virtual knockout of several receptor genes lowered exhaustion-programme connectivity, which prioritized targets only computationally.

    Significant positive GSEA enrichment of the NK-like programme and its receptor subset post-infusion vs infusion product; Monocle3 pseudotime rose from Tcm to Tem to Temra. Virtual KO of KLRC1, KLRD1 and KLRG1 reduced incoming signed connectivity of the exhaustion programme. · BioPathfinder-nominated patient dataset supports…; Virtual knockout

  4. ShownNKG2A was enriched on CD8+ CAR-T cells after chronic antigen exposure in vitro and in vivo, and NKG2A+ CD8+ CAR-T cells showed higher activation and exhaustion markers.

    Repeated Raji co-culture (fresh tumour cells every 4 days for 16 days) raised NKG2A+ frequency in CAR-T, most in CD8+. In NSG mice at day 20, NKG2A+ cells showed higher CD69, and significantly higher PD-1, LAG-3 and CTLA-4; TIGIT was not significant. · Chronic antigen stimulation enriches…; In vitro repeated antigen-stimulation assay of…; In vivo chronic stimulation assay of CAR-T cells

  5. SuggestedAdding anti-NKG2A antibody to CD19 CD8+ CAR-T cells improved tumour control, survival and effector readouts in a Raji-FG xenograft model, compared with CAR-T cells alone.

    Combination (monalizumab 10 mg/kg every 7 days) gave lower total bioluminescent flux, longer survival, higher serum IFN-γ, TNF-α and IL-2, and higher liver CAR-T frequency than CD8+ CAR-T alone. Group sizes and exact P values are not stated in the text. · NKG2A blockade enhances antitumour activity of…; In vivo NKG2A immune checkpoint blockade study…; Figure legends

Methods, data and reproducibility

A Curator agent extracts and adjudicates papers and accessions into a frozen evidence bundle; a Planner drafts hypothesis cards, LLM reviewers (baseline vs agent-assisted) score them, and a Bioinfo Analyst runs human-gated analyses before experimental testing of one expert-selected target.

  • Europe PMC/PubMed CAR-T single-cell/multiomic publication records dated 1 January 2017 to 4 May 2026; 3,889 retained after title/abstract prefilter · human (CAR-T-treated patient studies) · n = 16,597 records
  • Curated Curator corpus: dataset-linked patient scRNA-seq core studies plus supporting-only papers · human · n = 57 core studies; 10 supporting papers
  • GSE125881 patient CD19 CAR-T scRNA-seq across infusion product and post-infusion time points (patients CLL-1, CLL-2, NHL-6, NHL-7) · human · GSE125881
  • Planner-generated hypotheses scored for reviewer reproducibility · n = 50 hypotheses; three repeated runs
  • In vitro repeated Raji co-culture of healthy-donor CD19 CAR-T and non-transduced T cells · human
  • NSG mouse Raji and Raji-FG xenografts (chronic-stimulation phenotyping; NKG2A blockade with CD8+ CAR-T) · mouse (NSG) with human cells
Methods & tools
Europe PMC, GPT-5.4-mini, Scanpy, Scrublet, scVI-tools, CellTypist, edgeR, fgsea with MSigDB Hallmark, Monocle3, scTenifoldKnk, scBaseCount, CZ CELLxGENE
Software & environment
Code release v1.0.0-rc3; random seeds fixed where the tool exposes one; Immune_All_Low.pkl CellTypist model; FlowJo v.9; AURA imaging software 3.2.0. Software commit information is kept in the code release.
Assay conditions & compute
scVI trains on GPU when available, otherwise CPU; scTenifoldKnk default uses two CPU cores. No hardware models or runtimes are reported.
Reproducibility gaps
Patient scRNA-seq data are public under GEO accession GSE125881 (see methods.data); other data only on reasonable request to the corresponding author; file paths, schema identifiers and commit details are in the code release, not the Methods.
Limitations (6)
  • The evidence base covers only CAR-T-treated patient studies, which may exclude informative evidence from model systems, manufacturing studies and unpublished datasets. · stated by the authors
  • Hypothesis quality depends on the completeness of extracted metadata and evidence spans; unresolved paper-dataset links and incomplete sample annotations are hard to detect automatically. · stated by the authors
  • Bioinformatic analyses only support and prioritize; causal interpretation requires perturbation experiments. The system is not fully automatic, and humans scope analyses, resolve evidence links and pick targets. · stated by the authors
  • Generality of NK-like receptor perturbation needs prospective testing across more CAR constructs, tumour models and patient contexts. · stated by the authors
  • Reviewer evaluation measures score repeatability, not agreement with experimental outcomes; only one expert-selected hypothesis was tested, so the hit rate of prioritization is unknown. · noted in this summary
  • NKG2A blockade was shown in one Raji-FG xenograft with CD19 CD8+ CAR-T cells; the text gives no group sizes or exact P values, and no KLRC1 knockout is described. · noted in this summary

Read this if you want a template for turning curated patient single-cell evidence into agent-generated, expert-reviewed hypotheses with an experimentally tested CAR-T target.

Citations
1

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISMadan '25Daneshvar '26Moranzoni '26Luo '26Mergen '24Gordon '25
58% semantic matchBioinformatics
Sanna Madan, Tiangen Chang, Alexandra R. Harris+15 more·bioRxiv·Mar 19, 2025· 7 citations